Infection of the nematode C. elegans by pathogenic bacteria elicits learned pathogen avoidance behavior. Naïve animals dwell and feed on lawns of pathogenic P. aeruginosa (PA14), but after infection, animals avoid P. aeruginosa lawns (Meisel and Kim, 2014, Reddy et al., 2011). To obtain high-resolution measurements of learned pathogen avoidance behavior, we monitored animals that had been placed on a lawn of pathogenic PA14 for 18 hours. We counted the fraction of animals that remained on the lawn over time (Figure 1A) at 30-minute intervals. Within 8 hours, half of the animals changed their foraging behavior and left the lawn of pathogenic bacteria. By contrast, nearly all the animals placed on a control lawn of non-pathogenic Escherichia coli (OP50) bacteria remained on the lawn during 18 hours of monitoring (Figure 1A).

The fractional occupancy of worms on a bacterial lawn is affected in principle by two processes: lawn exit and lawn re-entry. We determined how PA14 exposure affected each of these processes. After four hours in the presence of pathogen, the rate constant for the lawn exit (exits per animal per hour) increased dramatically from zero and reached a plateau of roughly one exit event per worm-hour after 10 hours (Figure 1B). We observed that worms that left lawns of pathogen after 10–16 hours of exposure repeatedly attempted to return to the lawn but stalled for long periods upon contact with the lawn edge (Figure 1C, Video 1). We quantified this re-entry defect by measuring the latency to re-entry, that is, the delay between first contact with the lawn and re-entry of the animal into the lawn. Animals that had been exposed to a pathogen for short periods rapidly re-entered the lawn after an exit with latencies of only a few seconds (Figure 1D). After 5 hours of exposure to pathogen, the latency of re-entry began to increase, and after 12 hours of exposure, the mean latency to re-entry was 36.7±3.5 minutes (Figure 1D, Videos 2 and 3). Increases in lawn-leaving rates and latency to re-entry were correlated and coincided with the observed evacuation of the lawn of pathogenic bacteria (Figure 1D).

To demonstrate that the experience of pathogen exposure was driving the observed changes in behavior, as opposed to some change in the bacterial lawn caused by foraging animals, we performed an assay to test worm entry dynamics on a fresh lawn of PA14. We exposed animals to pathogenic PA14 for 15 hours, collected those that had left the lawn, and then compared the latency of lawn entry of this cohort to the latencies of a cohort of naive animals that had not experienced pathogen. Naive worms rapidly entered a fresh lawn of pathogen, whereas worms that had experienced 15 hours of pathogen exposure displayed increased latency of entry into such fresh lawns (Figure 1E–G). The results of this fresh lawn assay indicated that pathogen exposure changed the internal state of the worm to drive the observed behavioral transition from re-entering the lawn rapidly to stalled re-entry (Figure 1H).

We next sought to identify neurons that regulate interactions of C. elegans with lawns of pathogenic P. aeruginosa. Conventional approaches to identifying neurons required for a C. elegans behavior involve targeting individual neuron types by microablation (Avery and Horvitz, 1989, Gray et al., 2005, Kobayashi et al., 2013) or optogenetics (Liu et al., 2022a, López-Cruz et al., 2019, Sordillo and Bargmann, 2021). These approaches can be time-consuming, and applying them to perform unbiased screens of the entire nervous system can require the generation of large numbers of transgenic lines.

We previously showed that an optogenetic screen for neurons that drive a behavior can be performed more efficiently using multiplexed optogenetic manipulations of neurons followed by a compressed sensing analysis to infer individual key neuron types (Lee et al., 2019, Candes and Wakin, 2008, Donoho, 2006). We performed a compressed sensing-based screen using a panel of 29 transgenic C. elegans lines, each expressing the light-gated ion channel Archaerhodopsin-3 (Arch3) under a different promoter (Supplementary file 1). The panel used for this study focused on interneurons, and covered 54 classes of interneuron, 25 classes of sensory neuron, and 8 classes of motor neuron (Hobert et al., 2016; White et al., 1986; Xu et al., 2013). We optically inhibited neurons during the early stages of pathogen exposure when animals presumably associate pathogen-specific cues with sickness. We then monitored the subsequent dynamics of lawn leaving and re-entry. Animals were transferred to pathogenic lawns, given 1 hour to settle, and then illuminated for 2 hours with pulses of 525 nm green light (1 second on/off, 5 mW/mm²) to inhibit neurons expressing Arch3 in that line (Figure 2A). The behavior of these animals was then recorded for a total of 18 hours at 3-minute intervals (Figure 2A). We performed the same measurements of matched controls that had not been fed the opsin cofactor all-trans retinal and were thus insensitive to photoinhibition (Figure 2A). To quantify the effect of optogenetic inhibition on behavior, we calculated for each strain a differential retention index—the difference in the area between the temporal lawn occupancy curves of transgenic animals exposed to inhibitory light and a paired no-ATR control (Figure 2B). Six strains showed significant changes in differential retention index after neural silencing compared to controls (Figure 2B). Silencing neurons in these strains during the first 2 hours of pathogen exposure augmented lawn-leaving over the next 18 hours (Figure 1—figure supplement 1A).

We next asked whether the accelerated pathogen avoidance observed upon neural silencing resulted from increased lawn exits, increased latency to re-entry, or both. Two Archaerhodopsin lines (Pdop-2::Arch3 and Pmpz-1::Arch3) showed increased exit rates, while four lines (Pflp-4::Arch3, Psams-5::Arch3, Pttx-3::Arch3, and Pnpr-4::Arch3) did not show changes in exit rates upon inhibition (Figure 2C and D). Neural inhibition that increased lawn exit rates dramatically increased the number of tracks outside of the lawn (Figure 2E). We next measured the effects of the early neural silencing on lawn re-entry. We found that four lines (Pdop-2::Arch3, Pnpr-4::Arch3, Psams-5::Arch3, and Pttx-3::Arch3) showed significant increases in latency to re-entry in response to activation of Archaerhodopsin (Figure 2F and G). For example, the inhibition of the Pnpr-4::Arch3 line over the first 2 hours of the experiment dramatically decreased lawn re-entry over the entire time course, resulting in worm trajectories stalling at the lawn edge (Figure 2H). Inhibition of neurons on non-pathogenic OP50 produced dramatically different effects on both exit and entry compared to PA14 (Figure 1B and C). In particular, none of the optogenetic lines showed changes in latency to re-entry on OP50 (Figure 1C), suggesting that this effect is specific to PA14 exposure. These results indicated that inhibition of neurons during an early phase of learned pathogen avoidance could cause long-term changes in aversion to pathogenic bacteria through modulating distinct behaviors.

We next sought to identify specific neurons that influence learned pathogen avoidance. To identify neurons that control specific behaviors (Figure 3A), one usually thinks of perturbing N\begin{document}$N$\end{document} neurons of the nervous system one at a time. These measurements can be visualized as an N×N\begin{document}$N\times \, N $\end{document} set of equations, which can be easily solved to give the relative contribution of each neuron to the phenotype (Figure 3B). This approach requires as many measurements as the number of neurons being characterized. Utilizing a compressed sensing-based approach, we can instead formulate our optogenetic screen results as an underdetermined set of equations Mw→=P→.\begin{document}$\mathbf{M}\vec{w}=\vec{P}.$\end{document} The matrix, M\begin{document}$\mathbf{M}$\end{document}, is an incoherent measurement matrix of size 29 by 87 measurement matrix (Figure 3C), with each row corresponding to each Archaerhodopsin line (experiments were performed on 29 lines) and each column corresponding to neural identity. The matrix element, Mij\begin{document}$\mathbf{M}_{ij}$\end{document} is equal to 1 if the ith line drives expression in neuron j\begin{document}$j$\end{document} and is 0 otherwise. w⃑\begin{document}$\vec{w}$\end{document} corresponds to neural weights, that is, how much each neuron contributes to the phenotype, and P⃑\begin{document}$\vec{P}$\end{document} is the phenotype vector. The relative contributions (weights) of 87 neuron types to a phenotype can be determined using an L1 norm from these 29 measurements.

We evaluated the phenotype vector P⃑\begin{document}$\vec{P}$\end{document} for the two behaviors of interest by quantifying the changes to the dynamics of re-entry and exit from the lawns of pathogen. For both exit and re-entry, we calculated differences between optogenetically inhibited and non-inhibited worms for each transgenic line that showed statistically significant effects of inhibition. For lawn exit, the phenotype was the increase in the rate constant of exit due to neural inhibition. For re-entry, the phenotype was the difference between the latency in re-entry caused by neural inhibition. Lines that did not show significant behavioral changes in response to inhibition were assigned a phenotype value of zero.

We inferred neural weights w⃑\begin{document}$\vec{w}$\end{document} from our matrix equation using Lasso regression (Tibshirani, 2011, Tibshirani, 1996), which allowed us to solve the underdetermined set of linear equations Mw⃑=P⃑\begin{document}$\mathbf{M}\vec{w}=\vec{P}$\end{document} while simultaneously imposing sparsity constraints on the weight vector by minimizing the sum of the mean squared error χ2=(Mw→−P→)2\begin{document}$\chi ^{2}=\, \left (\mathbf{M}\vec{w}-\vec{P}\right)^{2}$\end{document} and the L1 norm of the solutions, thus minimizing Mw⃑-P⃑2+λw⃑1\begin{document}$\left (\mathbf{M}\vec{w}-\vec{P}\right)^{2}+\lambda \left \| \vec{w}\right \| _{1}$\end{document} where λ\begin{document}$\lambda $\end{document} is the sparsity parameter. Using this approach, we inferred a set of candidate neurons for lawn exit and lawn re-entry (Figure 3D and E). We focused primarily on neurons governing lawn re-entry (Figure 1F). For lawn re-entry, four key neurons were inferred over a wide range of sparsity parameters: AVK, SIA, AIY, and MI (Figure 2—figure supplement 1), with one example solution taken at a point where chi-squared error began to increase (Figure 3D). Some neurons were assigned negative weights by this analysis (suggesting that their inhibition promotes lawn re-entry). However, the contributions of these neurons decreased as the sparsity parameter increased, suggesting that these neurons were less important (Figure 2—figure supplement 1).

Compressed-sensing analysis of lawn-exit behavior identified contributions from six neuron classes to this behavior: CEPs, HSNs, RIAs, RIDs, and SIAs. Some of these neurons have previously been implicated in lawn retention or dwelling. RIAs are required for learned avoidance of Pseudomonas lawns (Liu et al., 2022b, Ha et al., 2010). CEPs and HSNs are also known to promote dwelling on bacterial lawns through release of dopamine and serotonin, respectively (Sawin et al., 2000). Notably, this analysis suggested that the neural circuit governing lawn exits is distinct from the neural circuit governing re-entry.

We next performed several tests to determine the robustness of our solutions, focusing on lawn re-entry behavior. To determine whether variation in archaerhodopsin expression might affect identification of neurons that govern this behavior, we tested the solutions to corrupted versions of the measurement matrix. The four neurons AVK, SIA, AIY, and MI were robustly identified regardless of matrix corruption (Figure 2—figure supplement 2). We next tested whether random removal of promoters would alter our solutions, i.e., whether a small subset of strains was driving the identification of neurons. Neuron identification was robust to removal of up to 5 of the 29 promoters (Figure 2—figure supplement 3). Finally, we determine false-positive and false-negative rates for compressed sensing-based inference (Figure 2—figure supplement 4A), as well as the recovery rate and true positive rate for each of the four neurons (AVK, SIA, AIY, and MI) (Figure 2—figure supplement 4B–E). While all four neurons identified have high recovery rates, MI, SIA, and AVK can have true recovery rates below 50%. Thus, we would expect at least one of these neurons to be a false positive. To directly test and determine the roles of the neurons implicated by compressed sensing, we next focused on measuring the activities from these neuron types in freely moving animals.

We measured calcium signals in three neuron subtypes (AVK, SIA, and AIY) in unrestrained worms before and after exposure to pathogenic Pseudomonas. We excluded MI from this analysis because of its role as a pharyngeal motor neuron; perturbation of MI might have affected lawn re-entry by modulating bacteria ingestion (Vidal et al., 2022).

To accurately measure calcium dynamics in AVK, SIA, and AIY, we used a custom real-time image-stabilization microscope capable of tracking and measuring neural dynamics in freely moving worms (Figure 4A). The microscope can track worms with 1 μm precision in all three dimensions while performing rotational stabilization by tracking a marker neuron (AWC^on, Pstr-2::mKO). We demonstrated the capability of this system by imaging neural activity in freely moving C. elegans at high magnification for up to an hour without affecting animal behavior (Lee et al., 2019).

We tracked neurons in transgenic animals expressing the calcium sensor GCaMP6s in AVK, SIA, or AIY interneurons. Naive animals (pre-pathogen exposure) were imaged on an empty agar plate for approximately 40 minutes. These worms were then placed onto a PA14 lawn for 24 hours, recovered to an empty assay plate, and imaged for approximately 40 minutes to determine how pathogen exposure affected neural activity. Pnpr-4::GCaMP6s was utilized to measure the activity of AVKs and SIAs, and Pttx-3::GCaMP6s was used to measure the activity of AIYs. We found that AIY (Figure 4C and D), AVK (Figure 4F and G), and SIA (Figure 4I and J) all showed reduced neural activity after pathogen exposure. These results were evident both in analyses of populations of animals (Figure 4C, F and I), but were also clearly observed in individual worms (Figure 4C, F and I inset and Figure 4D, G and J). Pathogen exposure had different effects on different interneurons. AIY neurons of naive animals showed a significant increase in calcium approximately 30 minutes after transfer to assay plates (Figure 4D). By contrast, AIY neurons of animals that had been exposed to pathogens remained quiescent. AVK and SIA neurons of naive animals displayed continuous high-frequency calcium signals. Post-exposure, AVKs and SIAs displayed long periods of quiescence. This data indicated that the activity of neurons that regulate a behavior critical for learned pathogen avoidance is strongly affected by exposure to pathogen. These neurons thus appear to encode the history of exposure to pathogenic bacteria in their neural activity state.

We next tested how manipulating the activity of neurons identified by compressed sensing as regulators of lawn re-entry affects how naïve animals interact with bacterial lawns. We inhibited each of the three candidate neuron types using Archaerhodopsin and measured how acute neuronal inhibitions affected re-entry into a lawn of pathogen. We performed parallel measurements of a control set of worms that had not been treated with the opsin cofactor ATR. To target specific neurons, we used a DLP mirror array to restrict illumination to cells of interest as previously described (Lee et al., 2019).

We found that acute inhibition of AIY in naive animals (no prior PA14 exposure) increased the latency of re-entry onto a fresh lawn of pathogenic Pseudomonas, mimicking the effect of prior pathogen exposure (Figure 5B). This effect was not the result of a general defect in locomotion or lawn-entry behavior; entry into lawns of non-pathogenic E. coli was not affected by AIY inhibition (Figure 5B). Inhibition of AVK failed to produce any effect on re-entry behavior (Figure 5D). Inhibition of npr-4-expressing neurons also increased latency to lawn entry on PA14 (Figure 5E). To validate that this effect was due to SIA, we projected patterned light onto Pnpr-4:Arch3 animals (Figure 5F) to selectively inhibit SIA/SIB. Inhibition of SIA using this method deterred entry of worms onto PA14 lawns (Figure 5G), resulting in increased latency in entry (Figure 5H). Overall, our results show that two out of the three key neurons identified by compressed sensing were able to elicit a change in lawn entry dynamics through inhibition. Together, our neural imaging and selective inhibition suggested that reduced activity of AIY, SIA, and SIB drives the reduction in lawn occupancy triggered by pathogen exposure.

We next tested whether activation of these neurons would suffice to reverse this behavioral switch. To test this, we optogenetically activated AIY, SIA, and SIB using transgenes expressing the light-gated ion channel channelrhodopsin-2 driven by the two promoters Pttx-3::ChR2 and Pnpr-4::ChR2. Transgenic lines were fed ATR and placed on lawns of pathogenic bacteria for 24 hours to induce lawn evacuation. Blue light 467 nm, 1 mW/mm² was then used to activate these neurons and the rate of re-entry onto the lawn was quantified. For both Pttx-3::ChR2 (Figure 5J) and Pnpr-4::ChR2 (Figure 5L), activation of these neurons dramatically increased the re-entry rate of experienced animals onto the PA14 lawn. Animals that re-entered the lawn also rapidly exited the lawn. Thus, the increased re-entry rate did not result in sustained increases in lawn occupancy (Figure 5—figure supplement 1). These results further illustrate that lawn exit and re-entry are controlled by a distinct set of neurons (Figure 2C and D) and demonstrate that both behavioral modules must change in order to evacuate the bacterial lawn.

All lines used in this work are listed in supplementary tables. Lines used for the measurement matrix are in Supplementary file 1. Lines used for neural activation and halorhodopsin can be seen in Supplementary file 2. Finally, lines used for neural imaging can be seen in Supplementary file 3. PDF-2 mutant used was wSR897: nlp-37(tm4780).

PA14 lawn evacuation assay plates were prepared as follows. PA14 was inoculated into LB media and allowed to grow for 15 hours without shaking at 37 °C until the culture reached an OD of 0.1–0.2. 5 uL of this culture was then pipetted onto NGM Agar plates and the colonies were allowed to grow for 18 hours at room temperature. Each colony was surrounded by a ring of filter paper to prevent worms from escaping to the edges of the plate. OP50 lawn evacuation plates were prepared in the same manner as PA14 evacuation plates; however, OP50 bacteria was grown for 18 hours at 37°C with shaking. This longer growth period was used to roughly match the thickness of the PA14 and OP50 colonies.

To begin lawn evacuation, 10 worms were transferred onto the bacteria lawn and given 1 hour to settle. All worms used for this assay were 1-day-old hermaphroditic adults grown on OP50 bacteria. The assay was then imaged at 3 fps for a total of 18 hours to assay lawn avoidance. Occupancy rate was calculated as (Number of worms on lawn)/(Total number of worms). Worms were counted as being on the lawn if any part of the body was in contact with the lawn.

PA14 bacteria colonies were generated as described for the lawn evacuation assay. Assays were initiated by placing 5–10 worms 5 mm away from the lawn edge, and then imaging the worms at 3 fps for 1 hour. Worms were infected prior to the assay by placing them onto a PA14 lawn evacuation colony for 15 hours. Worms that evacuated the colony over this time period were removed from the plate and used in the assay. Control worms were left on an OP50 lawn during this timeframe.

Compressed sensing was implemented using Lasso regression in Python using the scikit-learn library. Solutions were evaluated over a range of sparsity parameters over 6 orders of magnitude, and chi-squared error of the difference between behavioral predictions and measured behavior was calculated. Solutions shown in Figure 3 were taken by selecting solutions where chi-squared error began to increase as seen in Figure 2—figure supplement 1.

All transgenic lines used for this assay were generated by fusing Archaerhodopsin-3 to the relevant promoters via fusion PCR and then injecting the resulting constructs into worms. Worms used for optogenetic assays were fed on the rhodopsin cofactor all-trans retinal (ATR) for +12 hours before the assay. One-day-old adults were used for all behavioral assays.

Plates were set up as described for the lawn evacuation assays above. Following this, we performed optogenetic inhibition with pulsed (1 second on, 1 second off) 5 mW/mm² green light for 2 hours using a ScopeLED G250 to activate archaerhodopsin. Assays were imaged for a total of 18 hours either at one frame per 3 minutes for the full assay, or 3 fps to assay entry and exit sub-behaviors.

Targeted inhibition of SIA and SIB was carried out as described in previous work. Worms were cultured for 12+hours on ATR. Worms were then placed 5 mm away from the edge of a PA14 lawn generated as described above. SIA and SIB were located at the center of the fluorescent pattern Pnpr-4::Arch3 line. A circular pattern of light was projected from the DLP projector to selectively target SIA/SIB. Worms were tracked in the frame of view and imaged until they fully entered the PA14 colony or for a total of 1 hour after the worms first contacted the lawn.

Lawn exit rate was calculated by evaluating the number of exit events during the 2-hour timeframe of neural inhibition. Lawn evacuation movies were analyzed in 1-minute intervals, and any exit events occurring (as defined by a worm completely leaving contact with the bacteria) were noted. The exit events per hour were calculated from this, and statistics were calculated by aggregating exit rates per each experiment replicate (with each replicate consisting of 10 worms on a PA14 lawn). Latency to re-entry was also evaluated by measuring the time from first contact of the worm to the bacteria lawn to time of full entry of the worm onto the lawn. All latency to re-entry analysis was performed by aggregating all works that exited and re-entered the lawn during the measured timeframe. Analysis of statistical significance was performed by carrying out a t-test implemented in Matlab on the phenotype with and without neural inhibition.

Calcium imaging was performed using a custom-built microscope as previously described in past work at 15.625 frames per second. Worms were first imaged on an empty NGM plate for 40 minutes. Imaged worms were then transferred onto a colony of PA14 bacteria as used for the lawn evacuation assay and allowed to remain there for 24 hours. Following this, infected worms were imaged again for another 40 minutes.

GCaMP intensity information was extracted using custom software written in MATLAB to extract intensity data given segments encapsulating the neurons of interest. GCaMP imaging data was compiled for all imaged worms. Worm to worm variability in GCAMP expression was normalized by dividing all data by the bottom 5 percentile of fluorescence intensity in healthy worms. GCaMP data was smoothed over a 6-second window.

PA14 lawn evacuation plates were prepared as described above, and 10 ATR-treated worms expressing channelrhodopsin-2 under the relevant promoters were seeded onto each colony. Channelrhodopsin was activated using 1 mW/mm² blue light for 1 hour. Re-entry rate (defined as the rate at which worms fully entered the lawn) and contact rate were evaluated and quantified over this time.

The robustness of the solutions to removal of promoters was tested as follows. 1–5 promoters were removed from the measurement matrix at random. The resulting measurement matrix and phenotype vector were used to infer neuronal weights for the re-entry phenotype. This process was repeated 200 times for each of the 1–5 promoters, or 1000 times in total for 1–5 promoters. Our results demonstrated robustness of AIY, AVK, SIA, and MI to these removals.

The robustness of the solutions to corruption of the measurement matrix was tested by randomly altering a small fraction of the measurement matrix and re-inferring the neural contributions. 10% of the non-zero entries in the measurement matrix were altered to a value between 0 and 0.5. 1000 such corrupted measurement matrices were generated and solutions to each corrupted matrix were inferred via Lasso regression. Our results demonstrated that corruption generally did not result in alteration of the inferred neurons, with AIY, AVK, MI, and SIA robustly being inferred even with matrix corruption.

Single-cell RNA seq data from Taylor et al. was used to calculate the average expression of each neuropeptide within each neuron type. The relative expression of each of the neuropeptides in each neuron was calculated by dividing expression in each neuron by the maximum expression over all neurons. Inverse participation ratio (IPR) was calculated asIPR= 1∑i(Ei∑iEi)2\begin{document}$$\displaystyle IPR=\ \frac{1}{\underset{i}{\sum }\left (\frac{E_{i}}{\underset{i}{\sum }E_{i}}\right)^{2}}$$\end{document}

where Ei\begin{document}$E_{i}$\end{document} is the expression in the ith neuron.